Download Linear Continuous Maps and Topological Duals

c
Gabriel
Nagy
LCVS V
Locally Convex Vector Spaces V:
Linear Continuous Maps and Topological Duals
Notes from the Functional Analysis Course (Fall 07 - Spring 08)
In this section we take a closer look at continuity for linear maps from one locally convex
space into another.
A. Basic Definitions and Notations
Conventions & Notations. Throughout this note K will be one of the fields R or C,
and all vector spaces are over K. As a matter of terminology, given a vector space X , the
term functional on X designates a map φ : X → K.
Given two K-vector spaces X and Y, we denote by LinK (X , Y) the space of all K-linear
maps X → Y. (When there is no danger of confusion, the field K is omitted from the
notation.)
Given T ∈ Lin(X , Y) and x ∈ X , the vector T (x) will be denoted simply by T x. Given
S ∈ Lin(Y, Z), the composition S ◦ T ∈ Lin(X , Z) will be simply denoted by ST .
Suppose now T is a linear topology on X and S is a linear topology on Y. We denote
by LT,S
K (X , Y) the space of all K-linear operators, which are continuous with respect to the
given topologies. When there is no danger of confusion, the subscripts and superscripts will
be removed from the notation. In the case when (X , T) = (Y, S), the above space will
simply be denoted by LTK (X ), or simply L(X ).
A central problem in Functional Analysis is to identify topological vector spaces. For
this purpose we introduce the following terminology.
Definition. Suppose X1 and X2 are vector spaces, equipped with linear topologies T1
and T2 , respectively. A linear map T : X1 → X2 is said to be a topological linear isomorphism,
if T is bijective (hence a linear isomorphism), and both T : (X1 , T1 ) → (X2 , T2 ) and T −1 :
(X2 , T2 ) → (X1 , T1 ) are continuous.
B. Linear Continuous Functionals, Separation, and Closed Convex
Hulls
In the study of linear continuous maps, an important special case is always given great
attention, namely when the target space is the ground field K. If X is a vector space, the
space LinK (X , K), which is referred to as the algebraic dual of X , will be denoted simply by
X 0.
In the case when X is endowed with a linear topology T, and we endow K, with the
K
natural topology TK , the space LT,T
(X , K), is referred to as the (topological) T -dual space
K
of X , and is denoted simply by (X , T)∗ , or just X ∗ , when there is no danger of confusion.
1
The first result in this section is a collection of characterizations of linear continuous
functionals, most of which were already presented in earlier sections.
Proposition 1. Let X be vector space, equipped with a linear topology T.. For a linear
functional φ : X → K, the following are equivalent:
(i) φ is T-continuous1 ;
(i’) the functional Re φ : X → R is T-continuous;
(ii) φ is T-continuous at 0;
(ii’) the functional Re φ : X → R is T-continuous at 0;
(iii) Ker φ is T-closed;
(iii’) Ker Re φ is T-closed;
(iv) there exists s > 0, such that the set
Uφ (s) = {x ∈ X : |φ(x)| < s}
is a T-neighborhood of 0;
(iv’) there exists s > 0, such that the set
Xφ− (s) = {x ∈ X : Re φ(x) < s}
is a T-neighborhood of 0;
(v) for every s > 0 the set Uφ (s), defined in (iv), is a T-neighborhood of 0;
(v’) for every s > 0 the set Xφ− (s), defined in (iv’), is a T-neighborhood of 0;
(vi) |φ| is T-continuous.
Proof. The equivalences (i) ⇔ (ii) ⇔ (iii) and (i0 ) ⇔ (ii0 ) ⇔ (iii0 ) are shown in Lemma 2
from TVS III. The equivalence (i) ⇔ (i0 ) ⇔ (iv 0 ) ⇔ (v 0 ) is discussed in Lemm a 1 from CW
III. The implications (i) ⇒ (vi) ⇒ (v) ⇒ (ii) ⇒ (iv) are trivial. Finally, the implication
(iv) ⇒ (v) is easy to prove, since by linearity, for s, t > 0, the conditions |φ(x)| < s and
|φ(ts−1 x)| < t are equivalent, which shows that – using the notations from (iv) – one has
Uφ (t) = ts−1 Uφ (s), ∀ s, t > 0.
In particular, if Uφ (s) is a neighborhood of 0, then so is Uφ (t), for every t > 0.
1
It is understood that K is equipped with its natural topology.
2
Remark 1. Given a vector space X and a linear functional φ : X → K, the map
|φ| : X 3 x 7−→ |φ(x)| ∈ [0, ∞) defines a seminorm on X .
One important feature of topological duals in the locally convex Hausdorff case is described by the following result.
Proposition 2. If X is a locally convex topological vector space, then X ∗ separates the
points of X , in the following sense: for any x, y ∈ X , such that x 6= y, there exists φ ∈ X ∗ ,
such that φ(x) 6= φ(y).
Proof. Since X is locally convex and Hausdorff, there exists some open convex set A 3 y,
such that x 6∈ A. The existence of φ then follows from the “Easy” Hahn-Banach Separation
Theorem (from CW III).
Continuing our discussion on topological duals, we now take a closer look at an important
class of convex sets.
Definition. Suppose X is a vector space, equipped with a linear topology T. A subset
H ⊂ X is said to be a T-closed half-space in X , if there exist φ ∈ (X , T)∗ and s ∈ R, such
that
H = {x ∈ X : Re φ(x) ≤ s}.
Note that, by continuity, H is indeed closed. Note also that, by linearity, H is also convex.
Using this terminology, one has the following important result.
Theorem 1 (Closed Convex Hull Theorem) Suppose X is a locally convex topological
vector space. For any non-empty subset S ⊂ X , the closure conv(S) of the convex hull of
S is equal to the intersection of all closed half-spaces that contain S.
Proof. Let C denote the convex hull conv(S) and let C 0 denote the intersection of all closed
half-spaces that contain S. On the one hand, since all closed half-spaces are convex, it is
clear that C 0 is a convex set that contains S, thus C 0 contains C. On the other hand, since C 0
is also closed, being an intersection of closed sets, it follows that C 0 ⊃ C. To finish the proof,
we argue by contradiction, assuming the existence of some v ∈ C 0 , which does not belong
to C. By local convexity, there exists some open convex set A 3 v, such that A ∩ C = ∅.
By the Hahn-Banach separation Theorem (se CW III), there exists φ ∈ X ∗ and s ∈ R, such
that
Re φ(c) ≥ s > Re φ(a), ∀ a ∈ C, a ∈ A.
In particular, if we define H = {x ∈ X : Re φ(s) ≥ s}, then2 H is a closed-half-space which
contains C, but does not contain v. This forces v 6∈ C 0 , which is impossible.
Corollary 1. If X is a locally convex topological vector space, then for any linear subspace
Y, its closure is:
\
Y=
Ker φ.
(1)
∗
φ∈X
φ =0
Y
2
We can also write H = {x ∈ X : Re(−φ)(x) ≤ −s}.
3
Proof. Denote the right-hand side of (1) by Z. Clearly Z is a closed linear
subspace (since it
is an intersection of closed linear subspaces), which contains Y (since φY = 0 ⇔ Y ⊂ Ker φ),
therefore Z clearly contains Y. To prove the other inclusion, we must prove the implication
x 6∈ Y ⇒ x 6∈ X . Start with x 6∈ Y. By Theorem 1, there exists a closed half-space H that
contains Y, but not x. Write H = {x ∈ X : Re φ(x) ≤ s}, for some φ ∈ X ∗ and s ∈ R, so
that the conditions Y ⊂ H 63 x read:
Re φ(y) ≤ s < Re φ(x), ∀ y ∈ Y.
(2)
As it turns out, the first inequality in (2) in fact implies φY = 0. (Indeed, if this were not
s+1
true, one could find some y ∈ Y with φ(y) 6= 0, so the vector y1 = φ(y)
y ∈ Y would then
have φ(y
1 ) = s + 1, which is impossible.) Going back to (2) we now have s ≥ 0, so x 6∈ Ker φ.
Since φY = 0, this forces x 6∈ Z.
Comment. Theorem 1 raises an interesting problem. Assume T and T0 are two different
Hausdorff topologies on X . Then obviously one has
{A ⊂ X : A convex, T-open} =
6 {A ⊂ X : A convex, T0 -open}.
Despite the above inequality, is it possible to have the equality
{A ⊂ X : A convex, T-closed} = {A ⊂ X : A convex, T0 -closed}?
Surprisingly the answer is affirmative, as indicated by the following result. (See also the
example that follows, in which we actually have T0 ( T.)
Proposition 3. Given two locally convex Hausdorff topologies T and T0 on X , the
following statements are equivalent.
(i) For any convex set C, the conditions
• C is T-closed, and
• C is T0 -closed
are equivalent.
(ii) For any linear functional φ : X → K, the conditions
• φ is T-continuous, and
• φ is T0 -continuous
are equivalent, that is, one has the equality: (X , T)∗ = (X , T0 )∗ .
(iii) For any non-empty subset S ⊂ X , one has the equality
T
conv S = conv S
between T- and T0 -closures of conv S.
4
T0
(3)
Proof. (i) ⇒ (ii). Assume condition (i). To prove condition (ii) start with some linear
functional φ and we simply observe that, since Ker φ is obviously convex (in fact a linear
subspace), by (i) one has the equivalence
Ker φ T-closed ⇔ Ker φ T0 -closed
and then everything follows from Proposition 1.
(ii) ⇒ (iii). Assume (X , T)∗ = (X , T0 )∗ . In particular, for a linear functional φ and an
s ∈ R, when we consider the half-space H = {x ∈ X : Re φ(x) ≤ s}, the conditions
• H is T-closed, and
• H is T0 -closed
are equivalent. Therefore, for any S ⊂ X , one has the equality
{H ⊂ X , T-closed half-space, H ⊃ S} = {H ⊂ X , T0 -closed half-space, H ⊃ S},
(4)
and then (3) follows from Theorem 1.
The implication (iii) ⇒ (i) is trivial.
Example 1. Suppose (X , T) is a locally convex topological vector space. Consider the
collection Pw = {pφ : φ ∈ X ∗ } of seminorms, defined by pφ (x) = |φ(x)|. Consider the
locally convex topology wT , defined by Pw . Since X∗ separates the points, this topology
wT
is Hausdorff. Remark that, by construction, the condition xλ −−→
x is equivalent to the
∗
condition that pφ (xλ − x) → 0, ∀ φ ∈ X , which in turn is equivalent to:
φ(xλ ) → φ(x), ∀ φ ∈ X ∗ .
Therefore, wT is the weakest topology on X , with respect to which all φ ∈ X ∗ are continuous.
In particular, one has the inclusion wT ⊂ T, but also the equality of topological duals:
(X , wT )∗ = (X , T)∗ . In particular, by Proposition 3, it follows that, for every S ⊂ X , the TT
wT
and the wT -closure of conv S coincide: conv S = conv S .
Definition. The topology wT , constructed above, is called the weak topology implemented by T.
Example 2. Let X be an infinite dimensional vector space. Let T be the strongest
locally convex topology on X (see Exercise 1 from LCVS I). We know that T is defined
in such a way that every balanced convex absorbing set A is a T-neighborhood of 0. By
construction, all linear functionals φ ∈ X 0 are T-continuous. The weak topology wT is then
the locally convex topology described in Exercise 2 from LCVS I, so we know (see Exercise
3 from LCVS I) that wT ( T.
Comment. One general reason why the inclusion wT ⊂ T is strict in many instances,
is the following. If (X , T) is an infinite dimensional locally convex topological vector space,
the weak topology wT has a peculiar property: every wT -neighborhood of 0 contains a
closed infinite dimensional linear subspace. Indeed, if we start with some neighborhood V,
then using the notations from Example 1, there exist φ1 , . . . , φn ∈ X ∗ and ε1 , . . . , εn > 0,
5
such that V ⊃ ε1 B(pφ1 ) ∩ · · · ∩ εn B(pφn ), so V will clearly contain the closed subspace
(Ker φ1 ) ∩ · · · ∩ (Ker φn ).
C. Application: Compact Convex Sets
In this sub-section we discuss an important application of the Closed Convex Hull Theorem. In preparation for Theorem 2 below, we introduce the following terminology.
Definitions. Let X be a vector space, and let C ⊂ X be a non-empty convex subset. A
subset S ⊂ C is said to be extremal in C, if
(e) whenever x, y ∈ C are such that tx + (1 − x)y ∈ S, for some t ∈ (0, 1), it follows that
both x and y belong to S.
A point x ∈ C is called an extremal point in C, is the singleton set S = {x} is an extremal
subset in C.
Lemma 1. Suppose X is a locally convex topological vector space, C ⊂ X is a nonempty compact convex subset, and M ⊂ C is non-empty, compact, convex and extremal in
C. Suppose φ ∈ X ∗ , and define the quantity m = supx∈M Re φ(x). Then the set
S = {x ∈ C : Re φ(x) = m}
is non-empty, compact, convex, and extremal in C.
Proof. First of all, since M is compact, the maximum of (the continuous function) Re φ :
M → R is attained, i.e. S is non-empty.
Secondly, since φ is continuous, the set S is closed (in M), thus compact.
Thirdly, the convexity of S is quite obvious, from the linearity of φ.
Finally, to prove extremality, start with x, y ∈ C such that tx + (1 − t)y ∈ S, for some
t ∈ (0, 1), and let us prove that both x and y in fact belong to S. First of all, since
S ⊂ M, and M is extremal in C, it follows that both x and y belong to M. Argue now by
contradiction, assuming that either x 6∈ S or y 6∈ S. Then at least one of the inequalities
Re φ(x) ≤ m and Re φ(y) ≤ m would be strict, which (use the inequalities t > 0 and
1 − t > 0) forces
m = tm + (1 − t) > tRe φ(x) + (1 − t)Re φ(y) = Re φ(tx + (1 − t)y) = m
(last equality is due to the fact that tx+(1−t)y belongs to S), which is clearly impossible.
Theorem 2 (Krein-Milman). Suppose X is a locally convex topological vector space and
C ⊂ X is a non-empty compact convex subset.
(i) Every non-empty compact convex set A, which is an extremal subset of C, contains at
least on extreme point of C. In particular3 , the set
ext C = {x ∈ C : x extremal point in C}
is non-empty.
3
C is obviously extremal in itself.
6
(ii) conv(ext C) = C.
Proof. (i). Fix a non-empty compact convex extremal subset A of C. Consider the collection
E of all non-empty compact convex subsets of A, which are extremal subsets of C. Obviously
E is non-empty, since A ∈ E. Equip E with the order relation
E1 E2 ⇔ E1 ⊂ E2 .
We claim that E possesses at least one maximal element. To prove this, we apply Zorn
0
Lemma, so it suffices to show that any totally ordered
T sub-collection E = {Ei : i ∈ I}
admits an upper bound in E. Indeed, if we take E = i∈I Ei , then obviously E is compact
and convex. Since for every
T finite set of indices F ⊂ I, there is one j such that Ej ⊂ Ei ,
∀ i ∈ F , thus we have i∈F Ei ⊃ Ej 6= ∅, by compactness and the Finite Intersection
Property, it follows that E is also non-empty. Finally, the extremality of E is quite clear,
since if we start with two points x, y ∈ C and some t ∈ (0, 1), such that tx + (1 − t)y ∈ E,
then tx + (1 − t)y ∈ Ei , ∀ i ∈ I, so by the extremality of each Ei , we must have x, y ∈ Ei ,
∀ i ∈ I.
Fix now some maximal element M ∈ (E, ), and let us prove that M is a singleton.
Argue by contradiction, assuming M contains at least two points a 6= b. Choose then
a linear continuous functional φ ∈ X ∗ , such that φ(a) 6= φ(b). Multiplying, if necessary,
φ by −1 or ±i (in the complex case), we can assume Re φ(a) < Re φ(b). In particular,
if we define m = maxx∈M Re φ(x), we have the strict inequality Re φ(a) < m, so the set
S = {x ∈ M : Re φ(x) = m} does not contain a, which forces S ( M. We also know
by Lemma 1 that S is non-empty, compact, convex, and extremal in C, and this clearly
contradicts the maximality of M in (E, ).
(ii). Consider the set C0 = conv(ext C). Obviously, since C is convex, closed (in fact
compact), and it contains ext C, we have the inclusion C ⊃ C0 . To prove that in fact we have
the equality C = C0 , we argue by contradiction, assuming the existence of some a ∈ C which
does not belong to C0 . By the Closed Convex Hull Theorem, there exists φ ∈ X ∗ and some
s ∈ R, such that
Re φ(a) > s ≥ Re φ(x), ∀ x ∈ ext C.
(5)
Define, as in the proof of part (i), the quantity m = maxx∈C φ(x), and the set
S = {x ∈ C : φ(x) = m},
so that m > s, hence by (5) we have
S ∩ ext C = ∅.
(6)
By Lemma 1, S is non-empty, compact, convex, and extremal in C, so by (i) S must contain
an extreme point of C, which is impossible by (6).
In preparation for the next result (Theorem 3 below), the reader must solve the following:
Exercise 1. Let X be a locally convex topological vector space. Suppose C1 , . . . , Cn ⊂ X
are non-empty compact convex subsets, and let us consider the set C = conv(C1 ∪ · · · ∪ Cn ).
Prove that
7
(i) C is compact;
S
(ii) ext C ⊂ ni=1 ext Ci .
(Hint: Use induction, so everything is reduced to the case n = 2.)
Theorem 3 (Milman). Suppose X is a locally convex topological vector space and C ⊂ X
is a non-empty compact convex subset. If a subset S ⊂ X is such that
conv S = C,
then all extreme points of C belong to S.
Proof. Throughout the entire proof we can assume that K = R, that is, we are going to regard
X as a real locally convex topological vector space. (Convexity is, after all, dependent only
on the real vector space structure.)
Replacing S with its closure S, we can also assume that S is closed, hence compact (being
a subset of C). Fix some extreme point c ∈ ext C, and let us prove that c belongs to S.
Claim: For every finite set F ⊂ X ∗ and every ε > 0, there exists some point s ∈ S,
such that: |φ(s − c)| ≤ ε, ∀ φ ∈ F .
Replacing F with F ∪ (−1)F , suffices to prove the existence of s ∈ S, such that
φ(s − c) ≤ ε, ∀ φ ∈ F.
(7)
List now F = {φ1 , . . . , φn }, and define the sequence (Ck )nk=0 , (Dk )nk=0 recursively by C0 = C,
D0 = ∅, and
Ck = {x ∈ Ck−1 : φk (x − c) ≤ ε},
Dk = {x ∈ Ck−1 : φk (x − c) ≥ ε}.
Notice that, by construction, the Ck ’s and the Dk ’s are convex and closed, hence compact.
Since we have Ck−1 = Ck ∪ Dk we also have the equalities
C = Ck ∪ D1 ∪ · · · ∪ Dk , ∀ k = 1, . . . , n.
(8)
With these notations, we see that in order to prove the Claim, it suffices to show that
S ∩ Cn 6= ∅. Argue by contradiction, assuming that S is disjoint from Cn , which by (8),
would force S to be contained in the union M = D1 ∪ · · · ∪ Dn . In particular (by the above
Exercise), since M is a union of compact convex sets, its convex hull conv M is compact,
thus closed. Since S ⊂ M ⊂ C, it follows that
C = conv S ⊂ conv M ⊂ C,
so now we have C = conv(D1 ∪ · · · ∪ Dn ). By the above Exercise, the point c ∈ ext C must
belong to some Dj , which is clearly impossible, by the definition of the D’s.
Having proven the above Claim, it is now clear, that we can construct a net (sλ ) in S,
which converges to c in the weak topology wT , implemented by the given topology T on X .
Since S is T-compact, and wT is weaker than T, it follows that S is also wT -compact, hence
wT
also wT -closed. Now we are done, because the condition sλ −−→
c will force c ∈ S.
8
D. Linear Continuous Maps on Locally Convex Spaces
In the case of linear continuous maps between locally convex spaces, one can use the
metric point of view introduced in LCVS III, in conjunction with the following result.
Theorem 4. Suppose X and Y are locally convex spaces, whose topologies are defined
by two families of seminorms: P on X and Q on Y. For a linear map T : X → Y, the
following are equivalent.
(i) T is continuous.
(ii) T is continuous at 0.
(iii) For every q ∈ Q, there exist p1 , . . . , pn ∈ P, such that:
sup q(T x) : x ∈ B(p1 ) ∩ · · · ∩ B(pn ) < ∞.
(Recall that, for a seminorm p, the notation B(p) identifies the unit ball {x ∈ X :
p(x) < 1}.)
(iv) For every q ∈ Q, there exist p1 , . . . , pn ∈ P, and t1 , . . . , tn ≥ 0, such that:
q(T x) ≤ t1 p1 (x) + · · · + tn pn (x), ∀ x ∈ X .
(9)
Proof. The implication (i) ⇔ (ii) is trivial.
(ii) ⇒ (iii). Assume T is continuous at 0 and fix a seminorm q ∈ Q. Since q is continuous,
the unit ball B(q) = {y ∈ Y : q(y) < 1} is a (open) neighborhood of 0 in Y. In particular,
since T is continuous at 0, the preimage A = T −1 B(q), which is simply given by
A = {x ∈ X : q(T x) < 1}
is a neighborhood of 0 in X , so by construction (see LCVS III) there exist p1 , . . . , pn ∈ P
and ε1 , . . . , εn > 0, such that
ε1 B(p1 ) ∩ · · · ∩ εn B(pn ) ⊂ A.
Let ε = min{ε1 , . . . , εn }. Suppose now x ∈ B(p1 ) ∩ · · · ∩ B(pn ). Then
εx ∈ εB(p1 ) ∩ · · · ∩ εB(pn ) ⊂ ε1 B(p1 ) ∩ · · · ∩ εn B(pn ) ⊂ A,
so (by linearity) we get q(εT x) = q(T (εx)) < 1. In other words, we have
sup q(T x) : x ∈ B(p1 ) ∩ · · · ∩ B(pn ) ≤ ε−1 < ∞,
thus proving (iii).
(iii) ⇒ (iv). Assume condition (iii). Fix q ∈ Q, as well as p1 , . . . , pn ∈ P as in (iii), and
let t = sup q(T x) : x ∈ B(p1 ) ∩ · · · ∩ B(pn ) . To prove (iv), we now show that
q(T x) ≤ tp1 (x) + · · · + tpn (x), ∀ x ∈ X .
9
(10)
Start with some arbitrary x ∈ X , and then for every ε > 0, the vector
xε = [p1 (x) + · · · + pn (x) + ε]−1 x
satisfies
pk (xε ) =
pk (x)
< 1, ∀ k = 1, . . . , n,
p1 (x) + · · · + pn (x) + ε
so xε ∈ B(p1 ) ∩ · · · ∩ B(pn ) By condition (ii) it follows that q(T xε ) ≤ t, and then by linearity
we also get
q(T x) ≤ t[p1 (x) + · · · + pn (x) + ε].
Since the above inequality holds for all ε > 0, it forces the desired inequality (10)
(iv) ⇒ (i). Assume condition (iv), and let us prove that T is continuous. Start with
some net (xλ ) in X , which converges to some x ∈ X , and let us show that the net (T xλ )
converges to T x, which means:
q(T xλ − T x) → 0, ∀ q ∈ Q.
(11)
Fix q ∈ P and use condition (iv) to produce p1 , . . . , pn ∈ P and t1 , . . . , tn ≥ 0, satisfying
(9). One the one hand, by linearity, we have T xλ − T x = T (xλ − x), so (9) yields
0 ≤ q(T xλ − T x) ≤ t1 p1 (xλ − x) + · · · + tn pn (xλ − x).
(12)
On the other hand, since xλ → x, we also know that p(xλ − x) → 0, ∀ p ∈ P, so (12) clearly
forces q(T xλ − T x) → 0.
Comment. In the above Theorem, the equivalence (i) ⇔ (ii) holds in much greater
generality (see TVS I).
Remark 2. If in addition to the hypotheses in Theorem 2, we assume that P is directed,
then the four conditions are also equivalent to each of the following:
(iii’) For every q ∈ Q, there exists p ∈ P, such that:
sup q(T x) : x ∈ B(p) < ∞.
(iv’) For every q ∈ Q, there exist p ∈ P, and t ≥ 0, such that:
q(T x) ≤ tp(x), ∀ x ∈ X .
(13)
(First of all, since P is directed, for any p1 , . . . , pn ∈ P, there exist p ∈ P and s > 0, such
that p1 +· · ·+pn ≤ sp. In particular, condition (iv) from Theorem 2 clearly implies condition
(iv’) above. In turn, the inequality (13) clearly yields
sup{q(T x) : x ∈ B(p)} ≤ t < ∞,
so we also have the implication (iv 0 ) ⇒ (iii0 ). Finally condition (iii’) trivially implies condition (iii) from Theorem 2.)
Remarks 3-4. If we specialize to the case when Y = K, then the only seminorm needed
for K is q(y) = |y|. Therefore, in this case Theorem 2 and Remark 2 have the following
statements.
10
3. For a linear functional φ : X → K, the following are equivalent.
(i) φ is continuous.
(ii) φ is continuous at 0.
(iii) There exist p1 , . . . , pn ∈ P, such that:
sup |φ(x)| : x ∈ B(p1 ) ∩ · · · ∩ B(pn ) < ∞.
(iv) There exist p1 , . . . , pn ∈ P, and t1 , . . . , tn ≥ 0, such that:
|φ(x)| ≤ t1 p1 (x) + · · · + tn pn (x), ∀ x ∈ X .
4. If P is directed, then the above conditions are also equivalent to each of the following:
(iii’) There exists p ∈ P, such that:
sup |φ(x)| : x ∈ B(p) < ∞.
(iv’) There exist p ∈ P, and t ≥ 0, such that:
|φ(x)| ≤ tp(x), ∀ x ∈ X .
E. The Natural Topology on L(X , Y)
All spaces Lin(X , Y), L(X , Y), X 0 , and X ∗ are vector spaces, so it is legitimate to ask
whether they carry some natural linear topologies. As it turns out, to build such topologies all one needs is a linear topology on the target space Y, as indicated in the following
construction.
Definition. Suppose X and Y are vector spaces, and T is a linear topology on Y. Define,
for every x ∈ X the evaluation map
Ex : Lin(X , Y) 3 T 7−→ T x ∈ Y.
We can now consider the system E = (Ex )x∈X and apply the joint E-pull-back construction
from TVS II, to produce a topology denoted by Tso on Lin(X , Y). This topology is the
weakest topology on Lin(X , Y), which makes all the evaluation maps Ex , x ∈ X , continuous.
Since all Ex ’s are linear, Tso is a linear topology. In terms of convergence, for a net (Tλ ) and
Tso
some T in Lin(X , Y), the condition Tλ −−→
T is equivalent to:
T
Tλ x −
→ T x (in Y), ∀ x ∈ X .
The topology Tso is referred to4 as the strong operator topology induced by T. Note that, if
T is Hausdorff, then so is Tso (see TVS II).
4
There is nothing “strong” about this topology. We use this (unfortunate) terminology for consistency
reasons, in order to keep it in line with Hilbert space Operator Theory. It would be more natural to call it
the topology of point-wise convergence.
11
When we specialize to the case when Y = K (with its natural topology), so we work with
the algebraic dual space X 0 , this topology is simply denoted by w∗ , and is referred to as the
weak* (or weak dual ) topology. As noticed above, the weak* topology is always Hausdorff.
Convention. Assuming X is also equipped with a linear topology, we use the same
notations and terminology, when the above topologies are restricted to the spaces of linear
continuous maps on X . Thus L(X , Y) – which is a linear subspace in Lin(X , Y) – and X ∗
– which is a linear subspace in X 0 – will carry the induced topologies, which will be denoted
again by Tso and w∗ , respectively.
Remark 5. If Y is a complete topological vector space and X is an arbitrary vector
space, then (Lin(X , Y), Tso ) is also complete. Indeed, if we start with a net (Tλ )λ∈Λ in
Lin(X , Y), which is Tso -Cauchy, then for every x ∈ X , the net (Tλ x)λ∈Λ is Cauchy in Y, so
Tso
T.
it has a limit T x. It is pretty obvious that T is linear, and Tλ −−→
0
In particular (if we use Y = K), the algebraic dual X is complete, when equipped with
the w∗ -topology.
Remark 6. If the topology T (on Y) is locally convex, then the topology Tso on Lin(X , Y)
and on L(X , Y) is also locally convex. This follows from the discussion in LCVS II. Furthermore, if T is defined by a family Q of seminorms on Y, then the locally convex topology Tso
is defined by the family of seminorms
P = {q ◦ Ex : q ∈ Q, x ∈ X }.
In particular, the weak* topology on X 0 (or X ∗ ) is always locally convex, being defined by
the collection P = {px : x ∈ X }, where
px (φ) = |φ(x)|, ∀ φ ∈ X 0 , x ∈ X .
Exercises 2-3. Suppose Xi , i ∈ I are vector spaces. For any vector space Y, let
Y
Y
Xi )
Lin(Y, Xi ) → Lin(Y,
Π:
i∈I
i∈I
Σ:
Y
M
Lin(Xi , Y) → Lin(
Xi , Y)
i∈I
i∈I
be defined as follows:
Y
Π(T )y = (Ti y)i∈I , ∀ T = (Ti )i∈I ∈
Lin(Y, Xi ), y ∈ Y;
(14)
i∈I
Σ(S)x =
X
Si xi , ∀ S = (Si )i∈I ∈
i∈I
Y
Lin(Xi , Y), x = (xi )i∈I ∈
i∈I
M
Xi .
(15)
i∈I
(Note that in (15) the sum is finite.)
2. Prove that Π and Σ are linear isomorphisms.
3. Assume Y and Xi , i ∈ I, are locally convex spaces. Equip all Q
spaces L(Xi , Y) and
L(Y, Xi ), i ∈ I, with the strong operator topologies.
Let
X
=
prod
i∈I Xi be equipped
L
with the product topology, and let Xsum = i∈I Xi be equipped with the locally convex
sum topology. Prove the following.
12
Q
(i) If T = (Ti )i∈I ∈ i∈I L(Y,
Q Xi ), then Π(T ) : Y → Xprod is continuous. When we
equip the product space i∈I L(Y, Xi ) with the product topology, and the space
L(Y, Xprod ) with the strong operator topology, then the map
Y
Πc :
L(Y, Xi ) 3 T 7−→ Π(T ) ∈ L(Y, Xprod )
i∈I
is a topological linear isomorphism.
Q
(ii) If S = (Si )i∈I ∈ i∈I L(X
Qi , Y), then Σ(S) : Xsum → Y is continuous. When we
equip the product space i∈I L(Xi , Y) with the product topology, and the space
L(Xsum , Y) with the strong operator topology, then the map
Y
Σc :
L(Xi , Y) 3 S 7−→ Σ(S) ∈ L(Xsum , Y)
i∈I
is a topological linear isomorphism.
Comment. Assume Xi , i ∈ I, are locally convex. If we apply Exercise 3 with Y = K,
we obtain the existence of a natural topological linear isomorphism
Y
M ∗
∼
Σc :
(Xi )∗ −
→
Xi .
(16)
i∈I
i∈I
L
(It is understood that on Xsum =
convex sum topology, on
i∈I Xi one uses the locally
Q
all (Xi )∗ and Xsum one uses the weak* topology, and on i∈I (Xi )∗ one uses the product
topology.)
This is sometimes quoted as a basic rule by stating: “the dual of the sum is the product
of the duals.”
F. Alaoglu-Bourbaki Compactness Theorem
The result we are about to prove concerns compactness in the weak* topology. Since the
most popular applications come from a special case (for normed vector space version), we
will discuss them elsewhere.
Theorem 5 (Alaoglu-Bourbaki). Assume the vector space X is equipped with a linear
topology. For any neighborhood V of 0 in X , and any ρ > 0, the set
KV,ρ = {φ ∈ X ∗ : sup |φ(x)| ≤ ρ}
x∈V
∗
∗
is compact in (X , w ).
Proof. Consider the set T = {α ∈ K : |α| ≤ ρ}, and define the map
Y
Θ : KV,ρ 3 φ 7−→ φ(x) x∈V ∈
T.
V
Denote the range Θ(KV,ρ ) simply by R.
Q
Claim: When we equip KV,ρ with the w∗ -topology and we equip the set R ⊂ V T with
induced product topology, the map Θ : KV,ρ → R is a homeomorphism..
Remark first that Θ is injective. Indeed, if φ, ψ ∈ KV,ρ are such that Θ(φ) = Θ(ψ),
then φ(x) = ψ(x), ∀ x ∈ V, so since V is absorbing5 , by linearity, one also has φ(x) = ψ(x),
5
That is, for every x ∈ X , there exists t > 0 such that tx ∈ V
13
∀ x ∈ X , i.e. φ = ψ. To prove the fact that Θ is a homeomorphism, we must show that for
a net (φλ ) in KV,ρ and some φ ∈ KV,ρ , the conditions
w∗
• φλ −→ φ, and
• Θ(φλ ) → Θ(φ) (in
Q
V
T with respect to the product topology)
are equivalent. Explicitly this means that the conditions
φλ (x) → φ(x), ∀ x ∈ X ,
φλ (x) → φ(x), ∀ x ∈ V,
(17)
(18)
ought to be equivalent. This is, however, trivial, again using linearity and the fact that V is
absorbing.
Q
Remark now that, since T is compact, the product space V T is also compact in the
product topology, by Tihonov’s Theorem. Therefore, using the Claim, all we need to do is
to show that R is closedQin the product topology. Suppose (φλ ) is a net in KV,ρ , so that
Θ(φλ ) → Γ = (Γx )x∈V ∈ x∈V T in the product topology, and let us prove the existence of
(a unique) φ ∈ KV,ρ , such that Θ(φ) = Γ. Treating Γ as a function Γ : V → K, we know
that, by the definition of the product topology, we have
Γ(x) = lim φλ (x), ∀ x ∈ V.
(19)
λ
Remark now that:
(∗) if x ∈ X and α, β ∈ K are such that αx and βx both belong to V, then βΓ(αx) =
αΓ(βx).
This follows immediately from (19), which yields: βΓ(αx) = limλ φλ (αβx).
Using (∗) we can now define φ : X → K as follows. Start with some x ∈ X , choose some
t > 0 such that tx ∈ V (use here the fact that V is absorbing), and let φ(x) = t−1 Γ(tx). The
point of (∗) is that the value t−1 Γ(tx) is independent of the choice of t. By construction we
have
φ(x) = Γ(x), ∀ x ∈ V,
(20)
Property (∗) also yields:
φ(αx) = αφ(x), ∀ x ∈ X , α ∈ K.
To prove linearity, we start with x, y ∈ X and – using the fact that V is a neighborhood of
0 – we choose t > 0, such that tx, ty and tx + ty all belong to V. Then
φ(x + y) = t−1 Γ(tx + ty) = t−1 lim φλ (tx + ty) = t−1 lim[φλ (tx) + φλ (ty)] =
λ
λ
−1
−1
= t [lim φλ (tx) + lim φλ (ty)] = t [Γ(tx) + Γ(ty)] = φ(x) + φ(y).
λ
λ
Now we have a linear functional φ : X → K, satisfying
lim φλ (x) = φ(x), ∀ x ∈ V.
λ
14
(21)
Using the definition of KV,ρ , the limit (21) yields the inequality
|φ(x)| ≤ ρ, ∀ x ∈ V.
(22)
On the one hand, this shows that the set
Uφ (2ρ) = {x ∈ X : |φ(x)| < 2ρ}
contains V, so Uφ (2ρ) is a neighborhood of 0, and then by Proposition 1 it follows that φ is
indeed continuous. On the other hand, (22) now tells us that φ in fact belongs to KV,ρ , and
then (19) forces Γ = Θ(φ).
Remark 7. If Y is any linear subspace, such that X ∗ ⊂ Y ⊂ X 0 , then6 the set KV,ρ
is also compact in Y, in the induced w∗ -topology. In particular, KV,ρ is also compact in
(X 0 , w∗ ).
G. The (Topological) Transpose
Definitions. Suppose X and Y are vector spaces, and S ∈ Lin(X , Y), Then for any
vector space Z, one has a linear map
S transpZ : Lin(Y, Z) 3 L 7−→ LS ∈ Lin(X , Z).
The linear map S transpZ is referred to as the algebraic Z-transpose of S.
Assuming now X , Y, and Z are all equipped with linear topologies, and S is linear and
continuous, it is obvious that if L ∈ Lin(Y, Z) is continuous, then so is LS. Therefore, by
restricting S transpZ to L(Y, Z), we obtain a linear map
S ∗Z : L(Y, Z) 3 L 7−→ LS ∈ L(X , Z),
which is referred to as the topological Z-transpose of S. (When there is no danger of confusion, S ∗Z is simply denoted by S ∗ .
In particular, when Z = K, we speak of the topological dual transpose S ∗ : Y ∗ → X ∗ .
T
One key feature of the transpose operation is functoriality, which states that, if X −
→
S
M−
→ Y are linear and continuous, then the composition
S∗
T∗
L(Y, Z) −→ L(M, Z) −→ L(X , Z)
is (ST )∗ , i.e. one has the identity
(ST )∗ = T ∗ S ∗ .
Another key feature is continuity, as described in the following result.
Proposition 4. Let X , Y, and Z be equipped with linear topologies, and let S ∈ L(X , Y).
If we equip the spaces L(X , Z) and L(Y, Z) with the strong operator topologies, then the
linear map
S ∗ : (L(Y, Z), Ts-o ) → (L(X , Z), Ts-o )
is continuous.
In particular, when we equip the topological dual spaces X ∗ and Y ∗ with their weak*
topologies, the linear map S ∗ : (Y ∗ , w∗ ) → (X ∗ , w∗ ) is continuous.
6
This follows from a well know fact in Topology: if Y is a topological space, and Z is a subset in Y , then
every K ⊂ Z, which is compact in Z, relative to the induced topology, is also compact in Y .
15
T
T
so
so
S ∗ (L) in L(X , Z),
L in L(Y, Z), and let us prove that S ∗ (Lλ ) −−→
Proof. Suppose Lλ −−→
which means that
S ∗ (Lλ )x → S ∗ (L)x (in Z), ∀ x ∈ X .
The above condition is, however, trivial, since S ∗ (Lλ )x = Lλ (Sx) and S ∗ (L)x = L(Sx), and
Tso
L) that: Lλ y → Ly (in Z), ∀ y ∈ Y.
we know (by the condition Lλ −−→
Concerning the functoriality of the transpose operation, one has certain interesting connections, illustrated in the following two Exercises.
Exercise 4*. Suppose X is equipped with a linear topology, and Y is a locally convex
topological vector space. Prove that, for S ∈ L(X , Y), the following are equivalent:
(i) T has dense range in Y;
(ii) the topological dual transpose S ∗ : Y ∗ → X ∗ is injective;
(iii) for any vector space Z, equipped with a linear topology, the map S ∗ : L(Y, Z) →
L(X , Z) is injective.
(Hint: The implications (i) ⇒ (iii) ⇒ (ii) are trivial. When proving (ii) ⇒ (i), let M be
the closure of Range S, and argue
– using Corollary 1 – that, if M ( Y, then there exists
∗
y ∈ Y and φ ∈ Y , such that φM = 0, but φ(y) 6= 0.)
Exercise 5*. Let (Y, T) be a locally convex topological vector space, and let X ⊂ Y
be a linear subspace. Equip X with the induced topology TX , so that the inclusion map
J : X ,→ Y is continuous. Prove that the topological dual transpose J ∗ : Y ∗ → X ∗ is
surjective. (Hint: Let P be the directed collection of all T-continuous
seminorms on Y, so
that the induced topology TX is defined by the collection PX consisting of the restrictions
to X of all the seminorms in P. Argue that, if φ ∈ X ∗ , there exists p ∈ P, such that
|φ(x)
≤ p(x), ∀ x ∈ X . Use the Hahn-Banach Theorem to produce ψ ∈ Y ∗ , such that
ψ X = φ.)
Comment. When comparing the above two Exercises, one notices that in Exercise 4,
there is no counterpart of property (iii) from Exercise 4. In other words, using the notations
and assumptions as in Exercise 5, one cannot derive the surjectivity of
J ∗ : L(Y, Z) → L(X , Z),
(23)
for arbitrary (locally convex) topological vector spaces Z. It is quite trivial to see that the
surjectivity of (23) may even fail when Z = X . After all, the condition that
J ∗ : L(Y, X ) → L(X , X )
(24)
is surjective, is equivalent to the existence of some P ∈ L(Y, X ), satisfying
P x = x, ∀ x ∈ X .
A linear map P : Y → X , satisfying the above condition, is called a projection of Y onto
X . Using this language, the surjectivity of (23), for any Z, is equivalent to the existence
16
of a continuous projection P of Y ont X . (Indeed, if such a P exists, then J ∗ (LP ) = L,
∀ L ∈ L(X , Z).)
This discussion then legitimizes the formulation of the so-called Projection Problem:
Given a topological vector space Y, and a linear subspace X ( Y, decide if a continuous
projection of Y onto X exists. An equivalent of describing such projection is to think them
as operators Q ∈ L(Y), satisfying7 Q2 = Q and Range Q = X . It is trivial to see that
any Q ∈ L(Y), satisfying Q2 = Q, is forced to have closed range, therefore the Projection
Problem is only meaningful when X is a closed linear subspace. The two problems below
show examples when the Projection Problem has an affirmative answer. (A negative example
will be discussed in a different chapter.)
Exercise 6*. Assume Y is a locally convex topological vector space, and X ⊂ Y is a
finite dimensional linear subspace. Show that a continuous projection of Y onto X exists.
∼
(Hint: Fix a topological linear isomorphism S : X −
→ Kn , and consider the coordinate
maps πj : Kn → K, j = 1, . . . , n. Use Exercise 5 to obtain the existence of φ1 , . . . φn ∈ Y ∗ ,
such that (πj ◦ S)(x) = πj (x), ∀ x ∈ X , j = 1, . . . , n. Form the linear continuous operator
Φ = (φ1 , . . . , φn ) : Y → Kn , and let P = S −1 Φ.)
Exercise 7*. Assume Y is a locally convex topological vector space, and X ⊂ Y
is a closed linear subspace of finite co-dimension, i.e. the quotient space Y/X is finite
dimensional. Show that a continuous projection of Y onto X exists. (Hint: Let Π : Y →
Y/X denote the quotient map. Fix some linear basis {v1 , . . . , vn } for Y/X , let y1 , . . . , yn ∈ Y
be such that Πyj = vj , ∀ j = 1, . . . , n, and let S : Y/X → Y be the unique linear map
satisfying Svj = yj , ∀ j. Argue that Q : Y → Y, defined by Qy = y − SΠy, is the desired
projection.)
Exercise 8*. Suppose Y is a locally convex topological vector space, and X ⊂ Y
is a closed linear subspace. Using the notation from Exercise 5, we already know that
J ∗ : Y ∗ → X ∗ is surjective and continuous, when both X ∗ and Y ∗ are equipped with their
respective w∗ topologies. Prove that J ∗ is also open, so the w∗ topology on X ∗ is the quotient
topology on Y ∗ /Ker J ∗ . (Hint: Suffices to show that, if A is a basic open set in Y ∗ , which
can be written as A = {φ ∈ Y ∗ : maxy∈F |φ(y)| < 1}, for some finite subset F ⊂ Y, then
J ∗ A is open in X ∗ . Argue first that, if we consider the subspace Z = X + Span F, and the
inclusion K : Z ,→ Y, then K ∗ A is open in Z, in a very obvious way. This shows that we
can replace Y with Z, so we can take advantage of Exercise 7.)
Exercise 9. In Exercise 8, why do we need X to be closed in Y? Find an example,
where X ( Y is dense, such that J ∗ is not open.
Exercise 10. Let X be a locally convex topological vector space, and let Y ⊂ X
be a closed linear subspace. Equip the quotient space X /Y with the quotient topology,
and let Π : X → X /Y denote the quotient map. From Exercise 5 we already know that
Π∗ : L(X /Y, Z) → L(X , Z) is injective, for any topological vector space Z. Prove that
(i) Range Π∗ = {T ∈ L(X , Z) : T Y = 0};
(ii) When we equip Range Π∗ with the induced so topology, the map
Π∗ : (L(X /Y, Z), Tso ) → (Range Π∗ , Tso )
7
Using the multiplicative notation, Q2 stands for Q ◦ Q.
17
is a topological linear isomorphism.
Exercise 11. Suppose X , Y, and Z are vector spaces equipped with linear topologies,
and let S ∈ L(X , Y). Prove that the left composition operator
ΛS : (L(Z, X ), Tso ) 3 L 7−→ SL ∈ (L(Z, Y), Tso )
is continuous.
18